Fault diagnosis method, device and equipment of current transformer and storage medium

Abstract

The present application relates to the technical field of fault diagnosis, and discloses a current transformer fault diagnosis method, device, equipment and storage medium.The method comprises: current signal acquisition and complementary ensemble empirical mode decomposition are performed on the current transformer to obtain a plurality of intrinsic mode function components; time alignment and weighted reconstruction are performed on the intrinsic mode function components to obtain a current signal prediction sequence; Markov transition field conversion is performed on the current signal prediction sequence to generate a two-dimensional Markov transition field feature image; the two-dimensional Markov transition field feature image is input into a stripe pooling convolutional neural network, remote spatial features are extracted through a stripe pooling module, and then the feature channels are adaptively adjusted in weight through an SE attention mechanism module to obtain a fault type and a fault degree.The present application effectively solves the non-linear and non-stationary characteristics of signals and improves the fault diagnosis accuracy of the current transformer.

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis technology, and in particular to a fault diagnosis method, apparatus, equipment and storage medium for current transformers. Background Technology

[0002] Current transformers are crucial measuring devices in power systems, and their operating status directly impacts the safe and stable operation of the power grid. With the continuous expansion of the power grid and the deepening of smart grid construction, the difficulty of diagnosing current transformer faults is increasing. Traditional fault diagnosis methods mainly rely on manual experience and simple mathematical statistical analysis, which struggle to accurately identify fault types and severity under complex operating conditions, resulting in low diagnostic accuracy and poor reliability.

[0003] In recent years, with the development of artificial intelligence technology, deep learning-based fault diagnosis methods have been widely used. However, due to the high reliability and long-term stability of current transformers, obtaining fault samples is difficult, and the sample imbalance problem is serious, affecting the training effect of deep learning models. At the same time, the fault characteristics of current transformers under different operating conditions are complex and varied, making it difficult for traditional deep learning methods to effectively extract and utilize these features. Furthermore, the fault characteristics of current transformers are nonlinear and non-stationary, making it difficult for single signal processing methods to comprehensively capture fault information. Existing diagnostic methods often neglect the time-frequency characteristics and state transition features of signals, lacking the ability to quantitatively assess the degree of fault, which severely restricts the accuracy and practicality of fault diagnosis. Summary of the Invention

[0004] This invention provides a fault diagnosis method, apparatus, device, and storage medium for current transformers. This invention effectively solves the problems of nonlinearity and non-stationary characteristics of signals, and improves the accuracy of fault diagnosis for current transformers.

[0005] In a first aspect, the present invention provides a fault diagnosis method for a current transformer, the fault diagnosis method for the current transformer comprising:

[0006] Current signal acquisition and complementary ensemble empirical mode decomposition of current transformer are performed to obtain multiple intrinsic mode function components;

[0007] The intrinsic mode function components are time-aligned and weighted for reconstruction to obtain the current signal prediction sequence;

[0008] The current signal prediction sequence is subjected to Markov transfer field transformation to generate a two-dimensional Markov transfer field feature image.

[0009] The two-dimensional Markov transfer field feature image is input into a striped pooling convolutional neural network. The striped pooling module extracts the long-range spatial features, and the SE attention mechanism module adaptively adjusts the weights of the feature channels to obtain the fault type and fault degree.

[0010] Secondly, the present invention provides a fault diagnosis device for a current transformer, the fault diagnosis device for the current transformer comprising:

[0011] The acquisition module is used to acquire current signals from the current transformer and perform complementary ensemble empirical mode decomposition to obtain multiple intrinsic mode function components.

[0012] The reconstruction module is used to perform time alignment and weighted reconstruction on the intrinsic mode function components to obtain the current signal prediction sequence;

[0013] The conversion module is used to perform Markov transfer field conversion on the current signal prediction sequence to generate a two-dimensional Markov transfer field feature image.

[0014] The extraction module is used to input the two-dimensional Markov transfer field feature image into the striped pooling convolutional neural network, extract long-range spatial features through the striped pooling module, and then adaptively adjust the weights of the feature channels through the SE attention mechanism module to obtain the fault type and fault degree.

[0015] A third aspect of the present invention provides a fault diagnosis device for a current transformer, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor invokes the instructions in the memory to cause the fault diagnosis device for the current transformer to perform the aforementioned fault diagnosis method for the current transformer.

[0016] A fourth aspect of the present invention provides a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the above-described fault diagnosis method for a current transformer.

[0017] The technical solution provided by this invention effectively solves the problems of nonlinearity and non-stationary characteristics of the signal by introducing a complementary set empirical mode decomposition algorithm to adaptively decompose the original current signal, thereby improving the accuracy of feature extraction. An augmented Dickey-Fuller test and a dual-model prediction strategy are employed to specifically handle stationary and non-stationary components, enhancing the reliability of signal reconstruction. Markov transition fields are used for two-dimensional mapping of fault features, fully exploring the state transition characteristics of the signal and improving feature representation capabilities. A deep learning network structure with a striped pooling module and SE attention mechanism is designed to enhance the ability to capture long-range spatial features and highlight the weights of key feature channels. A fault diagnosis mechanism based on probability distribution entropy is established, enabling accurate determination of fault type and severity and providing reliable diagnostic basis. Through multi-level feature extraction and fusion strategies, the problem of insufficient samples is effectively overcome, improving the model's generalization ability. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 A flowchart illustrating the fault diagnosis method for a current transformer provided in this application embodiment;

[0020] Figure 2 A schematic block diagram of the structure of a fault diagnosis device for a current transformer provided in the embodiments of this application;

[0021] Figure 3 A schematic block diagram of the structure of a fault diagnosis device for a current transformer provided in an embodiment of this application. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] The flowchart shown in the attached diagram is for illustrative purposes only and does not necessarily include all content and operations / steps, nor does it necessarily have to be performed in the described order. For example, some operations / steps can be broken down, combined, or partially merged, so the actual execution order may change based on the actual situation.

[0024] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the scope of the application. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0025] It should also be further understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0026] The following detailed description of some embodiments of this application is provided in conjunction with the accompanying drawings. Unless otherwise specified, the following embodiments and features described herein can be combined with each other.

[0027] Please see Figure 1 , Figure 1 This is a flowchart illustrating the fault diagnosis method for a current transformer provided in an embodiment of this application, as shown below. Figure 1 As shown, the fault diagnosis method for current transformers provided in this application includes steps S100 to S600.

[0028] Step S100: Acquire current signals and perform complementary ensemble empirical mode decomposition on the current transformer to obtain multiple intrinsic mode function components;

[0029] It is understood that the executing entity of this invention can be a fault diagnosis device for a current transformer, or it can be a terminal or a server; no specific limitation is made here. This embodiment of the invention will be described using a server as an example.

[0030] Specifically, the raw current signal generated during the operation of the current transformer is acquired to obtain complete and high-precision raw data. Since the raw current signal is subject to pulse interference from the external environment during acquisition, median filtering is applied to effectively remove this pulse interference, resulting in a pulse-disrupted current signal. Outlier detection and removal are then performed on the pulse-disrupted current signal. Outliers in the current signal are caused by abnormal events or noise, which can affect the accuracy of subsequent analysis. Appropriate algorithms are used to detect and remove outliers, improving signal reliability. Wavelet transform is applied to the outlier-removed current signal, decomposing it into wavelet coefficients of different frequency bands. Wavelet transform effectively extracts local features of the signal, separating high-frequency noise from the signal's inherent details. Soft thresholding is then applied to the wavelet coefficients to suppress noise. Soft thresholding sets a threshold value; coefficients below this threshold are attenuated or even zeroed out, while coefficients above the threshold are retained, thus minimizing the impact of noise while preserving the main signal characteristics as much as possible. The denoised current signal is segmented, dividing the overall signal into several time segments so that each segment can be analyzed independently. The mean and variance of each segment are calculated, and the signal is standardized to eliminate biases introduced by different acquisition conditions or equipment characteristics, resulting in a standardized segmented current signal. The standardized segmented signals are then reconstructed into a complete standardized current signal through time-domain splicing. Complementary ensemble empirical mode decomposition (EMD) is performed on the standardized current signal. By introducing the complementary characteristics of white noise, the mode aliasing problem that easily occurs in traditional EMD is solved. During the decomposition process, the signal is decomposed layer by layer into several intrinsic mode function (IMF) components, each IMF component representing the characteristics of a specific frequency range of the signal.

[0031] Gaussian white noise with an amplitude of 0.2 times the standard deviation is added to the standardized current signal to generate N sets of auxiliary noise signals. The introduction of Gaussian white noise helps to solve the mode aliasing problem in traditional empirical mode decomposition, making the decomposition process more stable and robust. The auxiliary noise signals are superimposed on the standardized current signal to obtain a set of noisy signals. These noisy signals retain the main characteristics of the original signal, while the addition of different noises diversifies the signal. Local extrema are detected for each set of noisy signals to identify the maximum and minimum points of the signal. Cubic spline interpolation is performed on the maximum and minimum points to obtain the upper and lower envelopes. Cubic spline interpolation can accurately fit the envelope while ensuring smoothness, ensuring that it can closely fit the upper and lower boundaries of the signal. By calculating the average value of the upper and lower envelopes, a mean envelope is generated to reflect the local trend of the signal. The mean envelope is subtracted from the standardized current signal to obtain the candidate intrinsic mode functions. The candidate function is evaluated for its compliance with the intrinsic mode function (IMF) criteria using standard deviation calculation. An IMF must satisfy two conditions: first, the difference between the number of extreme points and zero-crossing points of the signal globally is at most 1; second, the mean envelope of the signal should be close to zero at any given time. If the candidate function does not meet these conditions, it is used as a new input signal, and the process of generating the envelope and calculating the mean envelope using cubic spline interpolation is repeated until the conditions are met. Once the conditions are met, the first IMF component is obtained. The first IMF component is subtracted from the normalized current signal to obtain the first decomposed residual signal. This residual signal is considered the new signal to be decomposed, and the above steps are repeated—from detecting local extreme points to extracting candidate IMFs—until the IMF conditions are met again, yielding the next IMF component. After each decomposition, the residual signal is updated to the new signal to be decomposed. This process is repeated until the residual signal becomes a monotonic function or its amplitude falls below a preset threshold. Under this termination condition, the decomposition process ends, yielding all intrinsic mode function (EMF) components and the final residual components. To enhance the accuracy of the decomposition, a ensemble average is performed on the EMF components obtained from the decomposition of N sets of auxiliary noise signals. By superimposing and averaging the results of multiple decompositions, random errors introduced by noise are eliminated, improving the stability and reliability of the decomposition results. After ensemble averaging, multiple EMF components are finally obtained.

[0032] Step S200: Time alignment and weighted reconstruction of the intrinsic mode function components to obtain the current signal prediction sequence;

[0033] Specifically, a preliminary analysis is performed on multiple intrinsic mode function (IMF) components, calculating their mean, variance, kurtosis coefficient, and skewness coefficient. These statistical characteristics describe the overall distribution characteristics of the signal. Simultaneously, frequency domain features are extracted to obtain indices such as spectral power and dominant frequency components, thus combining time-domain and frequency-domain features to form a time-frequency feature vector. Based on this vector, an augmented Dickey-Fuller test is performed on each IMF component to determine the signal's stationarity. The augmented Dickey-Fuller test calculates the test statistic and compares it with a critical value, dividing the IMF components into a set of stationary components and a set of non-stationary components. For the stationary component set, the signal's delay characteristics are analyzed using autocorrelation and partial autocorrelation functions, and the optimal order of the autoregressive moving average (ARMA) model is determined using the AIC criterion (Akaike Information Criterion). After determining the model order, the parameters of the ARMA model are solved using the least squares method, thereby achieving the modeling and prediction of the stationary components. Through this process, the future values ​​of the stationary components are estimated, forming the predicted values ​​of the stationary components. For a set of non-stationary components, the non-stationary components are input into a K-means clustering module, with the number of clusters set to k. Euclidean distance is used as the similarity metric to calculate the distance from each sample point to each cluster center, ultimately determining k cluster centers. Based on these cluster centers, the Gaussian kernel function values ​​of the non-stationary components are calculated, and these are used to construct the hidden layer nodes of a Radial Basis Function Neural Network (RBFNN). This method effectively captures the nonlinear characteristics of the non-stationary components. After the hidden layer nodes are constructed, orthogonal least squares are used to calculate the output matrix to obtain the weight parameters of the output layer. These weight parameters are then used to predict the non-stationary components, yielding their predicted values. The predicted values ​​of the stationary and non-stationary components are rearranged according to the original time series to ensure that the predicted sequence accurately corresponds to the time points in the original sequence, avoiding information misalignment. Due to the problem of unequal sampling points in the predicted values, cubic spline interpolation is used to process the sequence, achieving smooth alignment of irregular time intervals, resulting in an aligned predicted sequence. The aligned predicted sequence is then reconstructed using weighted energy components. The energy ratio of each component is calculated, and these ratios are used as weighting coefficients for sequence reconstruction. The purpose of weighted reconstruction is to adjust the predicted sequence according to the energy contribution of each component, making it closer to the characteristics of the original signal, and finally generating a current signal prediction sequence.

[0034] Step S300: Perform Markov transfer field transformation on the current signal prediction sequence to generate a two-dimensional Markov transfer field feature image;

[0035] Specifically, a time delay parameter τ is set for the current signal prediction sequence, where τ=1. The time delay parameter τ determines the reconstruction step size of the signal in phase space, and τ=1 is to maximize the preservation of the continuity of the time series. Based on this, the one-dimensional current signal prediction sequence is transformed into a reconstructed state matrix with a dimension of 10 using the phase space reconstruction method. This state matrix contains the embedded representation of the original time series in a multi-dimensional space. The state matrix is ​​discretized, transforming the continuous-valued state matrix into discrete-valued values, which facilitates statistical and probabilistic calculations. Each data point is uniformly divided according to the range of the maximum and minimum values ​​of the state matrix, and each data point is mapped to a predefined discrete state interval, resulting in a set of discrete state sequences. The discrete state sequences simplify the complex signal representation into a finite set of states. Statistical analysis is performed on adjacent state pairs in the discrete state sequences, calculating the number of times the state i transitions to state j, generating a state transition statistics matrix, where each element represents the transition frequency of a specific state pair. The state transition statistics matrix reflects the dynamic relationship between different states in the sequence. The transition statistics matrix is ​​normalized by dividing the transition frequency of each row by the total number of transitions in that row, resulting in the Markov transition probability matrix. Each element of the Markov transition probability matrix represents the conditional probability of transitioning from one state to another. The Markov transition probability matrix is ​​then normalized to ensure all probability values ​​are uniformly distributed within the range of 0 to 1, resulting in the normalized probability matrix. Based on this normalized probability matrix, a two-dimensional image matrix is ​​constructed, where each transition probability value is mapped to a corresponding pixel value, yielding the initial image of the Markov transition field. This image reflects the transition probability distribution of the signal between different states and is an important visual representation of time-series signals. Histogram equalization is then applied to the initial image of the Markov transition field. Histogram equalization redistributes the original pixel values ​​by calculating the cumulative distribution function of pixel values, thereby improving the overall contrast of the image. The equalized image better highlights hidden feature patterns in the signal. Finally, the enhanced feature image is smoothed using a 5×5 Gaussian kernel. Gaussian kernel smoothing effectively removes high-frequency noise from an image by calculating the weighted average of local pixels, while preserving the main feature structures, ultimately generating a two-dimensional Markov transfer field feature image.

[0036] Step S400: Input the two-dimensional Markov transfer field feature image into the striped pooling convolutional neural network, extract the long-range spatial features through the striped pooling module, and then adaptively adjust the weights of the feature channels through the SE attention mechanism module to obtain the fault type and fault degree.

[0037] Specifically, the 2D Markov transfer field feature image is input into the first convolutional layer of the striped pooling convolutional neural network. The first convolutional layer uses 32 3×3 kernels with a stride of 1 to extract initial features from the image through convolution operations. The convolutional feature map is then processed by the ReLU activation function; this nonlinear mapping enhances the network's ability to express features. Batch normalization is used to standardize the feature map, reducing the impact of internal covariate shifts and ensuring the consistency of feature distribution, resulting in the first layer feature map. This first layer feature map is then input into the striped pooling module of the striped pooling convolutional neural network. The striped pooling module uses 1×3 stripe kernels and employs a sliding window with a stride of 1 in the horizontal direction to aggregate features from the feature map. This design effectively captures the long-range dependencies of the feature map in the horizontal direction without compromising the integrity of features in the vertical direction. Max pooling is performed within the sliding window, retaining the feature values ​​with the strongest responses within the window to generate striped features. These striped features are then input into the second convolutional layer of the striped pooling convolutional neural network for feature extraction. In this layer, 64 3×3 convolutional kernels with a stride of 1 are used to extract high-level features of the image through convolution operations. The convolutional feature map is then non-linearly mapped using the ReLU activation function, enhancing the network's ability to capture complex feature patterns. Simultaneously, batch normalization is applied to adjust the distribution of the feature map, ensuring the stability and convergence of network training, resulting in the second layer feature map. This second layer feature map is then input into the squeezing layer of the SE attention mechanism module. The squeezing layer compresses the H×W dimension of the original feature map into a 1×1×C feature vector by performing global average pooling on each channel, generating channel descriptors. These channel descriptors are condensed representations of the global information of each channel, capturing the importance of features within each channel. The channel descriptors are then passed to the first fully connected layer of the SE attention mechanism module for dimensionality reduction. By reducing feature redundancy, more compact intermediate features are extracted. These intermediate features are then input into the second fully connected layer for dimensionality increase, restoring the feature dimension to the same number of channels as the original, and channel weight coefficients are generated using the Sigmoid activation function. These channel weight coefficients represent the importance of each channel and are used for subsequent feature recalibration. During the feature recalibration stage, each channel of the second-layer feature map is multiplied element-wise with its corresponding channel weight coefficient to enhance the feature response of important channels while suppressing the contribution of minor channels, generating fault feature mapping data. This fault feature mapping data combines the long-range spatial features extracted by the stripe pooling module with the channel-level weight information adaptively adjusted by the SE module, resulting in high representativeness and discriminative power. The probability distribution of each fault type is calculated based on the fault feature mapping data. Through the network's classification layer, the probability value corresponding to each fault type is extracted, and the uncertainty of the fault features is quantified by combining the entropy value of the probability distribution. The entropy value helps determine the severity of the fault; a high entropy value corresponds to a complex or severe fault, while a low entropy value indicates a simple or minor fault type.

[0038] The softmax function is used to perform a nonlinear transformation on each component of the fault feature mapping data. The softmax function transforms the input high-dimensional feature data into normalized probability values ​​between 0 and 1, while ensuring that the sum of all components is 1, resulting in a fault type probability vector with n fault types. Each component in this probability vector represents the normalized probability value of the corresponding fault type. A logarithmic transformation is performed on each component of the fault type probability vector to provide a nonlinear measure of the probability components for subsequent information entropy calculation, thus reflecting the uncertainty of the probability distribution of each type. According to the information entropy formula, the transformed components are weighted and summed to calculate the probability distribution entropy value of the diagnostic result. The information entropy value measures the uncertainty of the system; a higher entropy value indicates a more uniform probability distribution and higher uncertainty in the diagnostic result; while a lower entropy value indicates a more concentrated probability distribution and more reliable fault type diagnosis results. Based on the fault type probability vector, the fault type corresponding to the component with the highest probability value is selected as the candidate fault type. To ensure the reliability of the diagnostic result, this probability value is compared with a preset confidence threshold. If the probability value exceeds the confidence threshold, the preliminary diagnosis directly adopts the candidate fault type; otherwise, further analysis is needed to optimize the diagnosis. After the preliminary diagnosis is determined, the corresponding probability distribution entropy value is analyzed. If the probability distribution entropy value is found to be greater than the preset entropy threshold, this indicates that the system's diagnosis of this fault type has significant uncertainty, triggering a secondary diagnosis mechanism to further improve the accuracy of the diagnosis. In the secondary diagnosis mechanism, the fault type probability vector is matched and analyzed with a pre-stored fault feature library. Euclidean distance is used as a similarity metric to calculate the distance between the fault type probability vector and each sample in the feature library, and the feature template with the smallest distance is selected as the reference sample. By comparing historical data and typical features, the diagnosis result is corrected to obtain the corrected fault type. The probability distribution entropy value is mapped to intervals to quantify the severity of the fault. By defining a piecewise function, the entropy value is divided into three intervals, corresponding to mild, moderate, and severe fault degrees, respectively. Lower entropy values ​​correspond to minor faults with highly certain diagnostic results; medium entropy values ​​indicate some uncertainty in the system's diagnostic process, corresponding to moderate faults; while high entropy values ​​correspond to complex or severe fault types that are difficult for the system to classify clearly. Through this step, a complete diagnostic result of the fault type and severity is finally obtained.

[0039] In this embodiment of the invention, the original current signal is adaptively decomposed by introducing a complementary set empirical mode decomposition algorithm, effectively solving the problems of nonlinearity and non-stationary characteristics of the signal and improving the accuracy of feature extraction. An augmented Dickey-Fuller test and a dual-model prediction strategy are employed to specifically handle stationary and non-stationary components, enhancing the reliability of signal reconstruction. Markov transition fields are used for two-dimensional mapping of fault features, fully exploring the state transition characteristics of the signal and improving feature representation capabilities. A deep learning network structure with a striped pooling module and SE attention mechanism is designed to enhance the ability to capture long-range spatial features and highlight the weights of key feature channels. A fault diagnosis mechanism based on probability distribution entropy is established, enabling accurate determination of fault type and severity and providing reliable diagnostic basis. Through multi-level feature extraction and fusion strategies, the problem of insufficient samples is effectively overcome, improving the model's generalization ability.

[0040] In one specific embodiment, the process of performing step S100 may specifically include the following steps:

[0041] The current signal of the current transformer is acquired to obtain the raw current signal, and the raw current signal is then processed by median filtering to obtain a current signal with pulse interference removed.

[0042] Outlier detection and removal are performed on the current signal after removing pulse interference to obtain the current signal with outlier removed. Wavelet transform is then performed on the current signal with outlier removed to obtain wavelet coefficients.

[0043] The wavelet coefficients are subjected to soft thresholding to obtain the denoised current signal, and the denoised current signal is then segmented to obtain the segmented current signal.

[0044] The mean and variance of the segmented current signals are calculated to obtain the standardized segmented current signals, and the standardized segmented current signals are spliced ​​in the time domain to obtain the standardized current signals.

[0045] Complementary ensemble empirical mode decomposition is performed on the standardized current signal to obtain multiple intrinsic mode function components.

[0046] Specifically, the raw current signal of the current transformer is acquired in real time through sensors or acquisition devices. ,in This represents time. The original signal is affected by external environmental interference and contains non-stationary components such as high-frequency impulse noise, requiring median filtering. The mathematical expression for median filtering is:

[0047] ;

[0048] in, This is the signal after median filtering, with a window length of [value missing]. Median filtering effectively removes pulse interference while preserving the main signal trend. For signals with pulse interference removed... Outlier detection and removal are performed. Outliers are detected using statistical methods, such as outlier removal methods based on standard deviation.

[0049] ;

[0050] in, The mean of the signal. The standard deviation of the signal. This is the removal threshold (usually 2 or 3). After removing outliers, a new signal is obtained. Furthermore, its noise interference is further reduced. Wavelet transform is performed to extract multi-resolution features. The wavelet transform is expressed as:

[0051] ;

[0052] in, It is the scaled and translated form of the mother wavelet function. and These represent the scaling factor and the time shift factor, respectively. Wavelet transform decomposes a signal into wavelet coefficients of different frequency bands. This is used to analyze the local characteristics of a signal. To remove residual noise, the wavelet coefficients are soft-thresholded, and the formula is as follows:

[0053] ;

[0054] in, The processed wavelet coefficients, The threshold is used. Soft thresholding effectively reduces the impact of noise on the signal. For the denoised signal... Perform segmentation processing. Assume the signal is divided into segments... Segments, each segment is [length missing] Then the first The segment signal is represented as:

[0055] ;

[0056] For each signal segment, the mean and variance are calculated, and the standardization formula is as follows:

[0057] ;

[0058] in, and The first The mean and standard deviation of the segmented signal. After segmentation and standardization, all segmented signals are concatenated in the time domain to reconstruct the standardized current signal. For standardized signals Perform complementary set empirical mode decomposition. Add Gaussian white noise. :

[0059] ;

[0060] in, This represents the noise amplitude. Then, for... Perform empirical mode decomposition, decomposing it into several intrinsic mode function (IMF) components IM. By superimposing the decomposition results of multiple noise sets and averaging them, the final intrinsic mode function components are obtained:

[0061] ;

[0062] in, The number of noise groups, For the first The first group decomposition Each component. For example, if the original signal is a sinusoidal wave superimposed with noise, it can be decomposed into a pure low-frequency sinusoidal signal and a high-frequency noise component through complementary set empirical mode decomposition.

[0063] In one specific embodiment, the process of performing complementary ensemble empirical mode decomposition on the standardized current signal to obtain multiple intrinsic mode function components can specifically include the following steps:

[0064] M1: Add Gaussian white noise with an amplitude of 0.2 times the standard deviation to the standardized current signal to generate N sets of auxiliary noise signals;

[0065] M2: The auxiliary noise signal is superimposed with the standardized current signal to obtain a noisy signal group, and the local extrema of the noisy signal group are detected to obtain the maximum and minimum points of the signal.

[0066] M3: Perform cubic spline interpolation on the maximum and minimum points to obtain the upper and lower envelopes, and calculate the average value of the upper and lower envelopes to obtain the mean envelope;

[0067] M4: Subtract the mean envelope from the standardized current signal to obtain candidate intrinsic mode functions, and calculate the standard deviation of the candidate intrinsic mode functions to obtain the decomposed residuals;

[0068] M5: Based on the decomposition residual, determine whether the candidate intrinsic mode function meets the intrinsic mode function condition. If it does not meet the condition, repeat steps M3 to M4 until the condition is met, and obtain the first intrinsic mode function component.

[0069] M6: Subtract the first intrinsic mode function component from the normalized current signal to obtain the first decomposition residual signal, and repeat steps M2 to M5 with the first decomposition residual signal as the new signal to be decomposed to obtain the next intrinsic mode function component.

[0070] M7: Repeat steps M2 to M6 until the decomposed residual signal becomes a monotonic function or is less than the preset threshold, and obtain all intrinsic mode function components and the final residual component.

[0071] M8: Perform ensemble averaging on the intrinsic mode function components obtained from the decomposition of N sets of auxiliary noise signals to obtain multiple intrinsic mode function components.

[0072] Specifically, for standardized current signals Gaussian white noise is added to enhance signal stability. Assume the standard deviation of the normalized current signal is... The generated Gaussian white noise satisfy That is, the mean is zero and the standard deviation is 0.2 times the standard deviation of the original signal. This is achieved by superimposing the standardized current signal. The white noise generated independently by the group is obtained Group auxiliary noise signal Its mathematical expression is:

[0073] ;

[0074] in, Indicates the noise group number, The time variable is used. Each group of noisy signals... Perform local extremum detection to identify maxima in the signal. and minimum point This process is completed by determining the derivative, i.e., satisfying... And the second derivative A negative t represents a maximum, and a positive t represents a minimum. After detecting the extreme points, cubic spline interpolation is performed on these extreme points to generate the upper envelope of the signal. and lower envelope The interpolation function is expressed as:

[0075] ;

[0076] The mean envelope is obtained by calculating the average of the upper and lower envelopes. :

[0077] ;

[0078] Standardized current signal Subtract the mean envelope Candidate intrinsic mode functions are obtained. :

[0079] ;

[0080] For candidate intrinsic mode functions Calculate the standard deviation The decomposition residuals are obtained. :

[0081] ;

[0082] in, The total duration of the signal. yes The mean. If the residuals are decomposed... If the preset threshold is exceeded or the candidate intrinsic mode function does not meet the conditions for intrinsic mode functions (e.g., the difference between the number of zero crossovers and the number of extrema is no greater than 1, and the mean envelope is close to zero), then the envelope calculation and subtraction operation need to be re-executed until the conditions are met, and finally the first intrinsic mode function component is obtained. Subtract the standardized current signal The residual signal from the first decomposition was obtained. :

[0083] ;

[0084] by As a new signal to be decomposed, the process of extreme point detection, envelope calculation, subtraction operation, and intrinsic mode function condition judgment is repeated to sequentially extract subsequent intrinsic mode function components. This continues until the decomposed residual signal becomes a monotonic function or its amplitude is less than a preset threshold. For each group of noisy signals... After completing the above decomposition, the corresponding sets of intrinsic mode function components are obtained. By analyzing all... The ensemble average of the intrinsic mode function components of the noise signal group is used to obtain the final intrinsic mode function components:

[0085] ;

[0086] in, Indicates the first The first group of noise signal decomposition Each intrinsic mode function component.

[0087] In one specific embodiment, the process of performing step S200 may specifically include the following steps:

[0088] The mean, variance, kurtosis coefficient, and skewness coefficient of multiple intrinsic mode function components are calculated, and the frequency domain features of the intrinsic mode function components are extracted to obtain the time-frequency feature vector.

[0089] An augmented Dickey-Fuller test is performed on the intrinsic mode function components based on the time-frequency eigenvectors, and the intrinsic mode function components are divided into a set of stationary components and a set of non-stationary components according to the test statistic.

[0090] For the stationary component set, the optimal order of the autoregressive moving average model is determined by calculating the delay order of the autocorrelation function and the partial autocorrelation function, combined with the AIC criterion, and the model parameters are solved by the least squares method to obtain the predicted value of the stationary component.

[0091] Input the non-stationary component set into the K-means clustering module, set the number of clusters to k, and calculate the distance between the sample and the cluster center using Euclidean distance to obtain k cluster center points;

[0092] The Gaussian kernel function value is calculated based on the non-stationary component set of k cluster center points, and the hidden layer nodes of the radial basis function neural network are constructed to obtain the hidden layer output matrix.

[0093] The output layer weights are calculated using the orthogonal least squares method on the hidden layer output matrix, and the non-stationary components are predicted using the output layer weights to obtain the predicted values ​​of the non-stationary components.

[0094] The predicted values ​​of stationary components and non-stationary components are rearranged according to the original time series, and cubic spline interpolation is used to process the unequal interval sampling points to obtain the aligned prediction sequence.

[0095] The energy ratio of each component is calculated for the aligned prediction sequence. The energy ratio is then used as a weighting coefficient to reconstruct the sequence, resulting in the current signal prediction sequence.

[0096] Specifically, for the multiple eigenmode function components obtained from the decomposition... Perform basic statistical characteristic calculations, including the mean. ,variance kurtosis coefficient Skewing coefficient These characteristics reflect the basic statistical properties of the signal, and their calculation formulas are as follows:

[0097] ;

[0098] ;

[0099] in, The length of the time series. For the first The values ​​of each intrinsic mode function component. Frequency domain feature extraction is performed, and spectral characteristics are obtained through Fast Fourier Transform. Frequency domain features include the dominant frequency. and spectral power :

[0100] ;

[0101] in, for Fourier transform, This represents the upper frequency limit. Combining the above statistical and frequency domain characteristics, a time-frequency feature vector is constructed. Based on time-frequency feature vectors ,right Perform an augmented Dickey-Fuller (ADF) test to determine its stationarity. The null hypothesis of the ADF test is that the signal is non-stationary, and its test statistic is calculated using the following regression model:

[0102] ;

[0103] in, , For regression coefficients, Indicates the trend term. The coefficients for the delay order, This is the error term. When... When the t-statistic is significantly less than zero, the signal is considered stationary. Based on the test results, The model is divided into a set of stationary components and a set of non-stationary components. For the set of stationary components, the optimal order of the ARMA model is determined using the autocorrelation function (ACF) and the partial autocorrelation function (PACF). Selecting the optimal model order using the AIC criterion:

[0104] ;

[0105] in, These are the likelihood estimates of the model. After fitting the ARMA model, the least squares method is used to solve for the model parameters, yielding the predicted values ​​of the stationary components. For non-stationary component sets, input them into the K-means clustering module and set the number of clusters. Calculate samples using Euclidean distance With cluster center Distance:

[0106] ;

[0107] Select A number of cluster centers are identified, and a radial basis function neural network is constructed based on these centers. The Gaussian kernel function is calculated as follows:

[0108] ;

[0109] in, Define the kernel width. Construct the hidden layer output matrix. The output layer weights are calculated using the orthogonal least squares method. And predict the non-stationary components:

[0110] ;

[0111] The predicted values ​​of stationary and non-stationary components are rearranged according to the original time series order. Cubic spline interpolation is used to process the unequal interval sampling points to generate an aligned prediction sequence. The energy ratio of each component is calculated from the aligned prediction sequence. The formula is

[0112] ;

[0113] Will The weighting coefficients are used for sequence reconstruction to obtain the final predicted current signal sequence:

[0114] .

[0115] In one specific embodiment, the process of performing step S300 may specifically include the following steps:

[0116] Set a time delay parameter τ for the current signal prediction sequence, where τ=1, and perform phase space reconstruction on the current signal prediction sequence to obtain a state matrix with a reconstruction dimension of 10.

[0117] Discretize the state matrix by uniformly dividing each data point into intervals of maximum and minimum values ​​to obtain a discrete state sequence.

[0118] Statistical analysis is performed on the state pairs at adjacent time points in the discrete state sequence to calculate the number of times state i transitions to state j, and a state transition statistics matrix is ​​established.

[0119] Based on the state transition statistics matrix, the frequency of each state transition is normalized, and the frequency of each row is divided by the total number of transitions to obtain the Markov transition probability matrix.

[0120] The Markov transition probability matrix is ​​normalized to obtain a normalized probability matrix. A two-dimensional image matrix is ​​constructed based on the normalized probability matrix, and each transition probability value is used as the pixel value at the corresponding position to obtain the initial image of the Markov transition field.

[0121] Histogram equalization is performed on the initial image of the Markov transfer field. Pixel values ​​are redistributed by calculating the cumulative distribution function to obtain an enhanced feature image. Based on the enhanced feature image, a 5×5 Gaussian kernel smoothing process is performed to obtain a two-dimensional Markov transfer field feature image.

[0122] Specifically, the prediction sequence of the current signal Set time delay parameters ,make This allows for phase space reconstruction of the signal. Phase space reconstruction is a method that embeds a one-dimensional time series into a high-dimensional space to capture the dynamic characteristics of the signal. Assuming the reconstruction dimension is... The reconstructed state matrix Defined as:

[0123] ;

[0124] in, It is the total length of the sequence. It's a time delay. This is the embedding dimension. Each row represents a state vector, describing the local features of the signal at different time points. For the state matrix... Discretize the state by dividing its range into a finite number of discrete intervals. Assume the minimum and maximum values ​​of the state matrix are respectively... and Divide the value range evenly into There are 10 intervals, and the width of each interval is:

[0125] ;

[0126] For any element Map it to discrete states :

[0127] ;

[0128] in, , indicating the first The th state vector of the th state vector A series of discrete state values. By discretization, a continuous-value signal is represented as a discrete-state sequence. By statistically analyzing the state pairs at adjacent time points in the discrete state sequence, a state transition statistical matrix is ​​constructed. .set up Indicates from state Transition to state Frequency:

[0129] ;

[0130] After the statistics are completed, the frequency of each row is normalized to obtain the Markov transition probability matrix. :

[0131] ;

[0132] in, Indicates from state Transition to state The conditional probability satisfies Based on Markov transition probability matrix Probability normalization is performed to obtain the normalized probability matrix. Mapping all transition probability values ​​to the range [0,1] enhances the contrast of the probability distribution:

[0133] ;

[0134] Normalized probability matrix Mapped to a two-dimensional image matrix Each pixel value corresponds one-to-one with its normalized probability value.

[0135] ;

[0136] An initial image of the two-dimensional Markov transition field is generated, reflecting the transition probability distribution between states. Histogram equalization is applied to the initial image to enhance image features. The cumulative distribution function of pixel values ​​is calculated. , original pixel values Mapped to new pixel values :

[0137] ;

[0138] Histogram equalization enhances image contrast, which is beneficial for subsequent feature extraction. The enhanced image is then smoothed using a Gaussian kernel with a size of [missing value]. Gaussian kernel matrix :

[0139] ;

[0140] in, This is the standard deviation of the Gaussian kernel. Smoothing is done through convolution operations:

[0141] ;

[0142] The final generated two-dimensional Markov transfer field feature image It is smoother and clearer, and can effectively represent the dynamic behavior patterns of current signals.

[0143] In one specific embodiment, the process of performing step S400 may specifically include the following steps:

[0144] The two-dimensional Markov transfer field feature image is input into the first convolutional layer of the striped pooling convolutional neural network. Feature extraction is performed through 32 convolutional kernels of size 3×3 and convolution operation with stride of 1. The feature map is then processed by ReLU activation and batch normalization to obtain the first layer feature map.

[0145] The first layer feature map is input into the stripe pooling module of the stripe pooling convolutional neural network. A 1×3 stripe kernel and a horizontal sliding window with a stride of 1 are used for feature aggregation. Max pooling is performed on the features within the window to obtain the horizontal stripe features.

[0146] The stripe features are input into the second convolutional layer of the stripe pooling convolutional neural network for processing. 64 3×3 convolutional kernels and a stride of 1 are used to extract high-level features. ReLU activation and batch normalization are then performed to obtain the second layer feature map.

[0147] The second-layer feature map is input into the squeezing layer of the SE attention mechanism module. Global average pooling is performed on each feature channel to compress the H×W dimension feature map into a 1×1×C feature vector, thus obtaining the channel descriptor.

[0148] The channel descriptor is input into the first fully connected layer of the SE attention mechanism module for dimensionality reduction to obtain intermediate features, and the intermediate features are input into the second fully connected layer of the SE attention mechanism module for dimensionality increase to obtain channel weight coefficients.

[0149] The fault feature mapping data is obtained by multiplying each channel of the second-layer feature map element by element with the corresponding channel weight coefficient.

[0150] The probability distribution of each fault type is calculated based on the fault feature mapping data, and the fault type and fault severity are determined by combining the probability distribution entropy value.

[0151] Specifically, the two-dimensional Markov transfer field feature image The input is fed into the first convolutional layer of the striped pooling convolutional neural network. The first convolutional layer uses 32 arrays of size... convolution kernel With a stride of 1, local features are extracted from the input image through convolution. The mathematical expression of convolution is:

[0152] ;

[0153] in, Indicates the first Each convolutional kernel is located at... The output feature value, It's a bias. These are the weight parameters of the convolution kernel. Convolution operations can capture local feature patterns, such as edges and textures. The ReLU activation function is applied to the convolution result:

[0154] ;

[0155] The ReLU activation function introduces nonlinearity while suppressing negative values, thus improving the model's expressive power. The activation results are standardized using batch normalization, as shown in the formula:

[0156] ;

[0157] in, and These are the mean and variance of a small batch of data. and These are learnable parameters. This is a smoothing term to prevent the denominator from being zero. After convolution, activation, and normalization, the first layer feature map is obtained. The first layer feature map The input is fed into the stripe pooling module, which uses... The fringe kernel slides a window horizontally with a step size of 1, performing max pooling on the feature values ​​within each window. The formula for calculating fringe pooling is:

[0158] ;

[0159] This operation can aggregate horizontal feature information while maintaining the integrity of vertical features, generating horizontal stripe features. Stripe features The input is fed into the second convolutional layer. The second convolutional layer uses 64 elements of size [missing information]. convolution kernel With a stride of 1, higher-level feature extraction is performed. After the same convolution, ReLU activation, and batch normalization operations as the first layer, the second-layer feature map is obtained. The second layer feature map The input is fed into the squeeze layer of the SE attention mechanism module. The squeeze layer performs global average pooling on each feature channel. Spatial dimensions are compressed to generate channel descriptors. :

[0160] ;

[0161] in, It is a passage The channel descriptor z reflects the global importance of the channel. The channel descriptor z is input into the first fully connected layer of the SE module to complete the dimensionality reduction operation.

[0162] ;

[0163] in, and These are the weight matrix and bias of the first fully connected layer, respectively. It's an activation function. The intermediate features after dimensionality reduction. It is passed to the second fully connected layer for dimensionality upscaling, generating channel weight coefficients s:

[0164] ;

[0165] Here, the dimension of 's' is the same as the original number of channels. For the second layer feature map... Each channel is element-wise multiplied with its corresponding channel weight coefficient to achieve channel-level feature recalibration, thus obtaining fault feature mapping data. :

[0166] ;

[0167] Based on fault feature mapping data Calculate the probability distribution for each fault type. Map the features to the probability space using the softmax function:

[0168] ;

[0169] in, It is the first The probability of each fault type It is the classification weight. This represents the number of fault types. Combined with the entropy value of the probability distribution:

[0170] ;

[0171] Entropy This indicates the uncertainty of the prediction, and the severity of the fault can be further determined by comparing it with a set threshold. For example, when the entropy value is low and the probability of a certain fault type is significantly higher than that of other types, it is determined to be a fault of that type; while when the entropy value is high, it indicates that there is uncertainty in the diagnostic result.

[0172] In one specific embodiment, the process of calculating the probability distribution of each fault type based on the fault feature mapping data and determining the fault type and fault severity by combining the probability distribution entropy value can specifically include the following steps:

[0173] The fault feature mapping data is transformed nonlinearly by the softmax function, and the normalized probability values ​​of n fault types are calculated to obtain the fault type probability vector.

[0174] Logarithmic transformation is performed on each component in the fault type probability vector, and the transformed components are weighted and summed according to the information entropy formula to obtain the probability distribution entropy value of the diagnostic result.

[0175] Based on the fault type probability vector, the fault type with the highest probability value is selected as the candidate fault type, and the probability value is compared with the preset confidence threshold to obtain the preliminary diagnosis result;

[0176] The probability distribution entropy value corresponding to the preliminary diagnosis result is analyzed. When the probability distribution entropy value is greater than the preset entropy value threshold, the secondary diagnosis mechanism is triggered.

[0177] The Euclidean distance between the fault type probability vector and the pre-stored fault feature library is calculated, and the feature template with the smallest distance is selected as the reference sample to obtain the corrected fault type.

[0178] The probability distribution entropy value is mapped to three intervals using a piecewise function, corresponding to mild, moderate, and severe fault levels, respectively, thus obtaining the fault type and fault degree.

[0179] Specifically, map fault features to data Input the softmax function and its components Perform nonlinear transformation and calculate The normalized probability values ​​of each fault type are used to form a fault type probability vector. The mathematical expression for the softmax function is:

[0180] ;

[0181] in, It is the first The probability of each fault type These are the component values ​​corresponding to the fault characteristics. This represents the total number of fault types. The softmax function converts feature values ​​into probability values, ensuring that all... satisfy and For each component in the probability vector Perform a logarithmic transformation to obtain The probability distribution entropy value of the diagnostic results is calculated by weighting and summing the results according to the information entropy formula. The formula for information entropy is:

[0182] ;

[0183] in, It measures the uncertainty of the probability distribution. When When the value is close to zero, it indicates that the probability distribution is highly concentrated, and the system is highly certain about the diagnostic results; when... A larger value indicates a more dispersed distribution and greater uncertainty in the diagnostic results. (Based on the probability vector) The fault type with the highest probability value is selected as the candidate fault type. Let the maximum probability value be... The corresponding fault type is ,but:

[0184] ;

[0185] Will Compared with the preset confidence threshold In comparison, if If the preliminary diagnosis is accepted, then the diagnosis results will be adopted directly. As a diagnostic type; otherwise, further analysis of the corresponding entropy value is required. .when Greater than the preset entropy threshold When this occurs, it indicates high uncertainty in the diagnostic result, triggering a secondary diagnostic mechanism. In the secondary diagnostic mechanism, the fault type probability vector is... template vectors in the pre-stored fault feature library Euclidean distance is calculated, and the feature template with the smallest distance is selected as the reference sample to correct the fault type. The formula for Euclidean distance is:

[0186] ;

[0187] in, It is the first In the template vector, the th... The probability value of each component. Select... Make:

[0188] ;

[0189] Corresponding template This serves as the reference sample, and its type is used as the corrected fault type. Regarding the entropy value... Interval mapping is performed, dividing the fault into three intervals using a piecewise function, corresponding to mild, moderate, and severe fault levels, respectively. Let the entropy interval be... The mapping function is:

[0190] ;

[0191] By combining the corrected fault type and the fault severity corresponding to the entropy value, the final fault diagnosis result is generated.

[0192] Please see Figure 2 , Figure 2 A schematic block diagram of the structure of the fault diagnosis device 200 for a current transformer provided in this application embodiment, as shown below. Figure 2 As shown, the fault diagnosis device 200 for current transformers includes:

[0193] The acquisition module 210 is used to acquire current signals from the current transformer and perform complementary set empirical mode decomposition to obtain multiple intrinsic mode function components.

[0194] Reconstruction module 220 is used to perform time alignment and weighted reconstruction of intrinsic mode function components to obtain current signal prediction sequence;

[0195] The conversion module 230 is used to perform Markov transfer field conversion on the current signal prediction sequence to generate a two-dimensional Markov transfer field feature image.

[0196] The extraction module 240 is used to input the two-dimensional Markov transfer field feature image into the striped pooling convolutional neural network, extract the long-range spatial features through the striped pooling module, and then adaptively adjust the weights of the feature channels through the SE attention mechanism module to obtain the fault type and fault degree.

[0197] Through the synergistic collaboration of the aforementioned components, the original current signal is adaptively decomposed using a complementary set empirical mode decomposition algorithm, effectively addressing the nonlinear and non-stationary characteristics of the signal and improving the accuracy of feature extraction. An augmented Dickey-Fuller test and a dual-model prediction strategy are employed to specifically handle stationary and non-stationary components, enhancing the reliability of signal reconstruction. Markov transition fields are used for two-dimensional mapping of fault features, fully exploring the state transition characteristics of the signal and improving feature representation capabilities. A deep learning network structure with a striped pooling module and SE attention mechanism is designed, strengthening the ability to capture long-range spatial features and highlighting the weights of key feature channels. A fault diagnosis mechanism based on probability distribution entropy is established, enabling accurate determination of fault type and severity and providing reliable diagnostic basis. Through multi-level feature extraction and fusion strategies, the problem of insufficient samples is effectively overcome, improving the model's generalization ability.

[0198] Please see Figure 3 , Figure 3 This is a schematic block diagram of the structure of a fault diagnosis device 300 for a current transformer provided in an embodiment of this application. The fault diagnosis device 300 for a current transformer includes a processor 301 and a memory 302. The processor 301 and the memory 302 are connected through a device bus 303. The memory 302 may include a non-volatile storage medium and internal memory.

[0199] The non-volatile storage medium can store a computer program. The computer program includes program instructions, which, when executed by the processor 301, cause the processor 301 to perform any of the aforementioned fault diagnosis methods for current transformers.

[0200] The processor 301 provides computing and control capabilities to support the operation of the fault diagnosis device 300 for the entire current transformer.

[0201] The internal memory provides an environment for the execution of computer programs in non-volatile storage media. When the computer program is executed by the processor 301, the processor 301 can execute any of the above-mentioned fault diagnosis methods for current transformers.

[0202] Those skilled in the art will understand that Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the fault diagnosis device 300 of the current transformer involved in the present application. The specific fault diagnosis device 300 of the current transformer may include more or fewer components than shown in the figure, or combine certain components, or have different component arrangements.

[0203] It should be understood that processor 301 can be a Central Processing Unit (CPU), but it can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Among these, a general-purpose processor can be a microprocessor or any conventional processor.

[0204] It should be noted that, for the sake of convenience and brevity, the specific working process of the fault diagnosis device 300 for the current transformer described above can be referred to the corresponding process of the aforementioned fault diagnosis method for current transformers, and will not be repeated here.

[0205] This application also provides a computer-readable storage medium storing a computer program that, when executed by one or more processors, causes the one or more processors to implement the fault diagnosis method for a current transformer as provided in this application.

[0206] The computer-readable storage medium can be an internal storage unit of the fault diagnosis device 300 for the current transformer in the aforementioned embodiments, such as a hard disk or memory of the current transformer fault diagnosis device 300. Alternatively, the computer-readable storage medium can be an external storage device of the current transformer fault diagnosis device 300, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc., equipped with the current transformer fault diagnosis device 300.

[0207] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0208] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0209] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A fault diagnosis method of a current transformer, characterized by, The method comprises the following steps: current signal acquisition and complementary ensemble empirical mode decomposition are performed on the current transformer to obtain a plurality of intrinsic mode function components; time alignment and weighted reconstruction are performed on the intrinsic mode function components to obtain a current signal prediction sequence; specifically, mean, variance, kurtosis coefficient and skewness coefficient of the plurality of intrinsic mode function components are calculated, and frequency domain features of the intrinsic mode function components are extracted to obtain a time-frequency feature vector; an augmented Dickey-Fuller test is performed on the intrinsic mode function components based on the time-frequency feature vector, and the intrinsic mode function components are divided into a stationary component set and a non-stationary component set according to the test statistic; the stationary component set is processed by calculating the lag order of the autocorrelation function and the partial autocorrelation function, the optimal order of the autoregressive moving average model is determined by combining the AIC criterion, and the model parameters are solved by the least square method to obtain the prediction value of the stationary component; the non-stationary component set is input into a K-means clustering module, the number of clusters is set to k, the distance between the sample and the cluster center is calculated by the Euclidean distance to obtain k cluster center points; the non-stationary component set is calculated based on the k cluster center points to obtain the Gaussian kernel function value, the hidden layer nodes of the radial basis function neural network are constructed, and the hidden layer output matrix is obtained; the output layer weight is calculated by the orthogonal least square method, and the non-stationary component is predicted by the output layer weight to obtain the prediction value of the non-stationary component; the prediction value of the stationary component and the prediction value of the non-stationary component are rearranged according to the original time sequence, the unequal interval sampling points are processed by cubic spline interpolation, and the aligned prediction sequence is obtained; the energy ratio of each component is calculated, the energy ratio is used as a weighting coefficient for sequence reconstruction, and a current signal prediction sequence is obtained; the current signal prediction sequence is converted by Markov transition field to generate a two-dimensional Markov transition field feature image; the two-dimensional Markov transition field feature image is input into a stripe-pooling convolutional neural network, remote spatial features are extracted by a stripe-pooling module, and the feature channels are adaptively adjusted by an SE attention mechanism module to obtain the fault type and the fault degree.

2. The fault diagnostic method of a current transformer according to claim 1, characterized by, The method comprises the following steps: The current signal of the current transformer is collected to obtain an original current signal, and the original current signal is subjected to median filtering to obtain a current signal with pulse interference removed; outlier detection and rejection are performed on the current signal with pulse interference removed to obtain a current signal with outliers removed, and wavelet transform is performed on the current signal with outliers removed to obtain wavelet coefficients; soft threshold processing is performed on the wavelet coefficients to obtain a denoised current signal, and the denoised current signal is subjected to segmentation processing to obtain a segmented current signal; Perform mean and variance calculation on the segmented current signal to obtain a normalized segmented current signal, and perform time domain splicing on the normalized segmented current signal to obtain a normalized current signal; Perform complementary ensemble empirical mode decomposition on the normalized current signal to obtain a plurality of intrinsic mode function components.

3. The fault diagnostic method of a current transformer according to claim 2, characterized by, The complementary ensemble empirical mode decomposition of the normalized current signal to obtain a plurality of intrinsic mode function components comprises: M1: Add Gaussian white noise with an amplitude of 0.2 times the standard deviation to the normalized current signal to generate N groups of auxiliary noise signals; M2: Superimpose the auxiliary noise signals and the normalized current signal to obtain a noise signal group, and perform local extreme point detection on the noise signal group to obtain maximum and minimum points of the signal; M3: Perform cubic spline interpolation on the maximum and minimum points to obtain an upper envelope line and a lower envelope line, and calculate the average of the upper envelope line and the lower envelope line to obtain a mean envelope line; M4: Subtract the mean envelope line from the normalized current signal to obtain a candidate intrinsic mode function, and calculate the standard deviation of the candidate intrinsic mode function to obtain a decomposition residual; M5: Determine whether the candidate intrinsic mode function meets the intrinsic mode function condition based on the decomposition residual, and repeat steps M3 to M4 until the condition is met to obtain a first intrinsic mode function component; M6: Subtract the first intrinsic mode function component from the normalized current signal to obtain a first decomposition residual signal, and repeat steps M2 to M5 using the first decomposition residual signal as a new signal to be decomposed to obtain a next intrinsic mode function component; M7: Repeat steps M2 to M6 until the decomposition residual signal becomes a monotonic function or is less than a preset threshold to obtain all intrinsic mode function components and a final residual component; M8: Perform ensemble averaging on the intrinsic mode function components obtained by decomposing the N groups of auxiliary noise signals to obtain a plurality of intrinsic mode function components.

4. The fault diagnostic method of a current transformer according to claim 1, characterized by, The Markov transition field conversion of the current signal prediction sequence to generate a two-dimensional Markov transition field feature image comprises: Set a time delay parameter τ for the current signal prediction sequence, where τ=1, and perform phase space reconstruction on the current signal prediction sequence to obtain a state matrix with a reconstruction dimension of 10; Discretize the state matrix, and uniformly divide each data point according to the interval range of the maximum value and the minimum value to obtain a discrete state sequence; Statistically analyze the state pairs of adjacent time points in the discrete state sequence to calculate the number of times that state i is transferred to state j, and establish a state transition statistical matrix; Based on the state transition statistical matrix, normalize each state transition frequency by dividing the transition frequency of each row by the total transition frequency to obtain a Markov transition probability matrix; Perform probability normalization on the Markov transition probability matrix to obtain a normalized probability matrix, and construct a two-dimensional image matrix based on the normalized probability matrix, taking each transition probability value as a pixel value at the corresponding position to obtain a Markov transition field initial image; The initial image of the Markov transition field is subjected to histogram equalization, pixel values are redistributed by calculating the cumulative distribution function to obtain an enhanced feature image, and 5*5 Gaussian kernel smoothing processing is performed based on the enhanced feature image to obtain a two-dimensional Markov transition field feature image.

5. The fault diagnostic method of a current transformer according to claim 4, characterized by, The two-dimensional Markov transition field feature image is input into a stripe-pooling convolutional neural network, remote spatial features are extracted through a stripe-pooling module, and the feature channels are adaptively adjusted in weight through an SE attention mechanism module to obtain a fault type and a fault degree, including: The two-dimensional Markov transition field feature image is input into the first convolutional layer of the stripe-pooling convolutional neural network, feature extraction is performed through 32 convolution kernels with a size of 3*3 and a step of 1, and the feature map is subjected to ReLU activation and batch normalization processing to obtain a first layer feature map; The first layer feature map is input into the stripe-pooling module of the stripe-pooling convolutional neural network, feature aggregation is performed using a stripe kernel with a size of 1*3 and a horizontal sliding window with a step of 1, and a maximum value pooling operation is performed on the features in the window to obtain horizontal stripe features; The stripe features are input into the second convolutional layer of the stripe-pooling convolutional neural network for processing, high-level features are extracted using 64 convolution kernels with a size of 3*3 and a step of 1, and ReLU activation and batch normalization processing are performed to obtain a second layer feature map; The second layer feature map is input into the squeeze layer of the SE attention mechanism module, a global average pooling operation is performed on each feature channel, the feature map with a dimension of H*W is compressed into a feature vector with a dimension of 1*1*C to obtain a channel descriptor; The channel descriptor is input into the first fully connected layer of the SE attention mechanism module for dimension reduction to obtain intermediate features, and the intermediate features are input into the second fully connected layer of the SE attention mechanism module for dimension increase to obtain channel weight coefficients; Each channel of the second layer feature map is multiplied with the corresponding channel weight coefficient element by element to perform channel-level feature re-labeling to obtain fault feature mapping data. The probability distribution of each fault type is calculated according to the fault feature mapping data, and the fault type and the fault degree are determined in combination with the probability distribution entropy value.

6. The fault diagnostic method of a current transformer according to claim 5, characterized by, The probability distribution of each fault type is calculated according to the fault feature mapping data, and the fault type and the fault degree are determined in combination with the probability distribution entropy value, including: The components of the fault feature mapping data are subjected to nonlinear transformation through a softmax function to calculate the normalized probability values of n fault types to obtain a fault type probability vector; Each component in the fault type probability vector is subjected to logarithmic transformation, and the weighted sum of the transformed components is calculated according to the information entropy formula to obtain the probability distribution entropy value of the diagnosis result; Based on the fault type probability vector, the fault type with the maximum probability value is selected as the candidate fault type, and the probability value is compared with the pre-set confidence threshold to obtain a preliminary diagnosis result; The probability distribution entropy value corresponding to the preliminary diagnosis result is analyzed, and when the probability distribution entropy value is greater than a preset entropy value threshold, a secondary diagnosis mechanism is triggered; The fault type probability vector is calculated with the Euclidean distance of the pre-stored fault feature library, the smallest distance feature template is selected as the reference sample, and the corrected fault type is obtained; The probability distribution entropy value is interval mapped, the entropy value is divided into three intervals through a segmented function, corresponding to mild, moderate and severe fault degrees, and the fault type and fault degree are obtained.

7. A fault diagnostic device for a current transformer, characterized by comprising: A fault diagnosis method for a current transformer as claimed in any one of claims 1-6, comprising: A collection module for collecting current signals of the current transformer and performing complementary ensemble empirical mode decomposition to obtain a plurality of intrinsic mode function components; A reconstruction module for time aligning and weighting reconstructing the intrinsic mode function components to obtain a current signal prediction sequence; A conversion module for converting the current signal prediction sequence through a Markov transition field to generate a two-dimensional Markov transition field feature image; An extraction module for inputting the two-dimensional Markov transition field feature image into a stripe pooling convolutional neural network, extracting remote spatial features through a stripe pooling module, and then adaptively adjusting the feature channels through an SE attention mechanism module to obtain the fault type and fault degree.

8. A fault diagnostic device for a current transformer, characterized by comprising: The fault diagnosis device of the current transformer comprises a memory and at least one processor, and the memory stores instructions; The at least one processor calls the instructions in the memory to enable the fault diagnosis device of the current transformer to perform the fault diagnosis method of the current transformer as claimed in any one of claims 1-6.

9. A computer-readable storage medium having stored thereon instructions, the computer-readable storage medium comprising: The instructions are executed by the processor to implement the fault diagnosis method of the current transformer as claimed in any one of claims 1-6.

Images